Uni-Fold: Training your own deep protein-folding models.

Related tags

Deep LearningUni-Fold
Overview

Uni-Fold: Training your own deep protein-folding models.

This package provides and implementation of a trainable, Transformer-based deep protein folding model. We modified the open-source code of DeepMind AlphaFold v2.0 and provided code to train the model from scratch. See the reference and the repository of DeepMind AlphaFold v2.0. To train your own Uni-Fold models, please follow the steps below:

1. Install the environment.

Run the following code to install the dependencies of Uni-Fold:

  conda create -n unifold python=3.8.10 -y
  conda activate unifold
  ./install_dependencies.sh

Uni-Fold has been tested for Python 3.8.10, CUDA 11.1 and OpenMPI 4.1.1. We recommend using Conda >= 4.10 when installing the environment: using Conda with lower level may lead to some conflicts between packages.

2. Prepare data before training.

Before you start to train your own folding models, you shall prepare the features and labels of the training proteins. Features of proteins mainly include the amino acid sequence, MSAs and templates of proteins. These messages should be contained in a pickle file /features.pkl for each training protein. Uni-Fold provides scripts to process input FASTA files, relying on several external databases and tools. Labels are CIF files containing the structures of the proteins.

2.1 Datasets and external tools.

Uni-Fold adopts the same data processing pipeline as AlphaFold2. We kept the scripts of downloading corresponding databases for searching sequence homologies and templates in the AlphaFold2 repo. Use the command

  bash scripts/download_all_data.sh /path/to/database/directory

to download all required databases of Uni-Fold.

If you successfully installed the Conda environment in Section 1, external tools of search homogenous sequences and templates should be installed properly. As an alternative, you can customize the parameters of feature preparation script to refer to your own databases and tools.

2.2 Run the preparation code.

An example command of running the feature preparation pipeline would be

  python generate_pkl_features.py \
    --fasta_dir ./example_data/fasta \
    --output_dir ./out \
    --data_dir /path/to/database/directory \
    --num_workers 1

This command automatically processes all FASTA files under fasta_dir, and dumps the results to output_dir. Note that each FASTA file should contain only one sequence. The default number of cpu used in hhblits and jackhmmer are 4 and 8. You can modify them in unifold/data/tools/hhblits.py and unifold/data/tools/jackhmmer.py, respectively.

2.3 Organize your training data.

Uni-Fold uses the class DataSystem to automatically sample and load the training entries. To make everything goes right, you shall pay attention to how the training data is organized. Two directories should be established, one with input features (features.pkl files, referred as features_dir) and the other with labels (*.cif files, referred as mmcif_dir). The feature directory should have its files named as _ _ /features.pkl , and the label directory should have its files named as .cif . Users shall make sure that all proteins used for training have their corresponding labels. See ./example_data/features and ./example_data/mmcif for instances of features_dir and mmcif_dir.

3. Train Uni-Fold.

3.1 Configuration.

Before you conduct any actual training processes, please make sure that you correctly configured the code. Modify the training configurations in unifold/train/train_config.py. We annotated the default configurations to reproduce AlphaFold in the script. Specifically, modify the data setups in unifold/train/train_config.py:

"data": {
  "train": {
    "features_dir": "where/training/protein/features/are/stored/",
    "mmcif_dir": "where/training/mmcif/files/are/stored/",
    "sample_weights": "which/specifies/proteins/for/training.json"
  },
  "eval": {
    "features_dir": "where/validation/protein/features/are/stored/",
    "mmcif_dir": "where/validation/mmcif/files/are/stored/",
    "sample_weights": "which/specifies/proteins/for/training.json"
  }
}

The specified data should be contained in two folders, namely a features_dir and a mmcif_dir. Organizations of the two directories are introduced in Section 2.3. Meanwhile, if you want to specify the subset of training data under the directories, or assign customized sample weights for each protein, write a json file and feed its path to sample_weights. This is optional, as you can leave it as None (and the program will attempt to use all entries under features_dir with uniform weights). The json file should be a dictionary contains the basename of directories of protein features ([pdb_id]_[model_id]_[chain_id]) and the sample weight of each protein in the training process (integer or float), such as:

{"1am9_1_C": 82, "1amp_1_A": 291, "1aoj_1_A": 60, "1aoz_1_A": 552}

or for uniform sampling, simply using a list of protein entries suffices:

["1am9_1_C", "1amp_1_A", "1aoj_1_A", "1aoz_1_A"]

Meanwhile, the configurations of models can be edited in unifold/model/config.py for users who want to customize their own folding models.

3.2 Run the training code!

To train the model on a single node without MPI, run

python train.py

You can also train the model using MPI (or workload managers that supports MPI, such as PBS or Slurm) by running:

mpirun -n <numer_of_gpus> python train.py

In either way, make sure you properly configurate the option use_mpi in unifold/train/train_config.py.

4. Inference with trained models.

4.1 Inference from features.pkl.

We provide the run_from_pkl.py script to support inferencing protein structures from features.pkl inputs. A demo command would be

python run_from_pkl.py \
  --pickle_dir ./example_data/features \
  --model_names model_2 \
  --model_paths /path/to/model_2.npz \
  --output_dir ./out

or

python run_from_pkl.py \
  --pickle_paths ./example_data/features/1ak0_1_A/features.pkl \
  --model_names model_2 \
  --model_paths /path/to/model_2.npz \
  --output_dir ./out

The command will generate structures of input features from different input models (in PDB format), the running time of each component, and corresponding residue-wise confidence score (predicted LDDT, or pLDDT).

4.2 Inference from FASTA files.

Essentially, inferencing the structures from given FASTA files includes two steps, i.e. generating the pickled features and predicting structures from them. We provided a script, run_from_fasta.py, as a more friendly user interface. An example usage would be

python run_from_pkl.py \
  --fasta_paths ./example_data/fasta/1ak0_1_A.fasta \
  --model_names model_2 \
  --model_paths /path/to/model_2.npz \
  --data_dir /path/to/database/directory
  --output_dir ./out

4.3 Generate MSA with MMseqs2.

It may take hours and much memory to generate MSA for sequences,especially for long sequences. In this condition, MMseqs2 may be a more efficient way. It can be used in the following way after it is installed:

# download and build database
mkdir mmseqs_db && cd mmseqs_db
wget http://wwwuser.gwdg.de/~compbiol/colabfold/uniref30_2103.tar.gz
wget http://wwwuser.gwdg.de/~compbiol/colabfold/colabfold_envdb_202108.tar.gz
tar xzvf uniref30_2103.tar.gz
tar xzvf colabfold_envdb_202108.tar.gz
mmseqs tsv2exprofiledb uniref30_2103 uniref30_2103_db
mmseqs tsv2exprofiledb colabfold_envdb_202108 colabfold_envdb_202108_db
mmseqs createindex uniref30_2103_db tmp
mmseqs createindex colabfold_envdb_202108_db tmp
cd ..

# MSA search
./scripts/colabfold_search.sh mmseqs "query.fasta" "mmseqs_db/" "result/" "uniref30_2103_db" "" "colabfold_envdb_202108_db" "1" "0" "1"

5. Changes from AlphaFold to Uni-Fold.

  • We implemented classes and methods for training and inference pipelines by adding scripts under unifold/train and unifold/inference.
  • We added scripts for installing the environment, training and inferencing.
  • Files under unifold/common, unifold/data and unifold/relax are minimally altered for re-structuring the repository.
  • Files under unifold/model are moderately altered to allow mixed-precision training.
  • We removed unused scripts in training AlphaFold model.

6. License and disclaimer.

6.1 Uni-Fold code license.

Copyright 2021 Beijing DP Technology Co., Ltd.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0.

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

6.2 Use of third-party software.

Use of the third-party software, libraries or code may be governed by separate terms and conditions or license provisions. Your use of the third-party software, libraries or code is subject to any such terms and you should check that you can comply with any applicable restrictions or terms and conditions before use.

6.3 Contributing to Uni-Fold.

Uni-Fold is an ongoing project. Our target is to design better protein folding models and to apply them in real scenarios. We welcome the community to join us in developing the repository together, including but not limited to 1) reports and fixes of bugs,2) new features and 3) better interfaces. Please refer to CONTRIBUTING.md for more information.

Owner
DeepModeling
Define the future of scientific computing together
DeepModeling
Learning Temporal Consistency for Low Light Video Enhancement from Single Images (CVPR2021)

StableLLVE This is a Pytorch implementation of "Learning Temporal Consistency for Low Light Video Enhancement from Single Images" in CVPR 2021, by Fan

99 Dec 19, 2022
Implementation and replication of ProGen, Language Modeling for Protein Generation, in Jax

ProGen - (wip) Implementation and replication of ProGen, Language Modeling for Protein Generation, in Pytorch and Jax (the weights will be made easily

Phil Wang 71 Dec 01, 2022
Code for SIMMC 2.0: A Task-oriented Dialog Dataset for Immersive Multimodal Conversations

The Second Situated Interactive MultiModal Conversations (SIMMC 2.0) Challenge 2021 Welcome to the Second Situated Interactive Multimodal Conversation

Facebook Research 81 Nov 22, 2022
CMSC320 - Introduction to Data Science - Fall 2021

CMSC320 - Introduction to Data Science - Fall 2021 Instructors: Elias Jonatan Gonzalez and José Manuel Calderón Trilla Lectures: MW 3:30-4:45 & 5:00-6

Introduction to Data Science 6 Sep 12, 2022
Source Code and data for my paper titled Linguistic Knowledge in Data Augmentation for Natural Language Processing: An Example on Chinese Question Matching

Description The source code and data for my paper titled Linguistic Knowledge in Data Augmentation for Natural Language Processing: An Example on Chin

Zhengxiang Wang 3 Jun 28, 2022
This is an official implementation for "SimMIM: A Simple Framework for Masked Image Modeling".

SimMIM By Zhenda Xie*, Zheng Zhang*, Yue Cao*, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai and Han Hu*. This repo is the official implementation of

Microsoft 674 Dec 26, 2022
NEG loss implemented in pytorch

Pytorch Negative Sampling Loss Negative Sampling Loss implemented in PyTorch. Usage neg_loss = NEG_loss(num_classes, embedding_size) optimizer =

Daniil Gavrilov 123 Sep 13, 2022
The final project for "Applying AI to Wearable Device Data" course from "AI for Healthcare" - Udacity.

Motion Compensated Pulse Rate Estimation Overview This project has 2 main parts. Develop a Pulse Rate Algorithm on the given training data. Then Test

Omar Laham 2 Oct 25, 2022
Cluster-GCN: An Efficient Algorithm for Training Deep and Large Graph Convolutional Networks

Cluster-GCN: An Efficient Algorithm for Training Deep and Large Graph Convolutional Networks This repository contains a TensorFlow implementation of "

Jingwei Zheng 5 Jan 08, 2023
Joint project of the duo Hacker Ninjas

Project Smoothie Společný projekt dua Hacker Ninjas. První pokus o hříčku po třech týdnech učení se programování. Jakub Kolář e:\

Jakub Kolář 2 Jan 07, 2022
Predict multi paths to a moving person depending on his trajectory history.

Multi-future Trajectory Prediction The project is about using the Multiverse model to make possible multible-future trajectory prediction for a seen p

Said Gamal 1 Jan 18, 2022
Link prediction using Multiple Order Local Information (MOLI)

Understanding the network formation pattern for better link prediction Authors: [e

Wu Lab 0 Oct 18, 2021
A self-supervised 3D representation learning framework named viewpoint bottleneck.

Pointly-supervised 3D Scene Parsing with Viewpoint Bottleneck Paper Created by Liyi Luo, Beiwen Tian, Hao Zhao and Guyue Zhou from Institute for AI In

63 Aug 11, 2022
Unbalanced Feature Transport for Exemplar-based Image Translation (CVPR 2021)

UNITE and UNITE+ Unbalanced Feature Transport for Exemplar-based Image Translation (CVPR 2021) Unbalanced Intrinsic Feature Transport for Exemplar-bas

Fangneng Zhan 183 Nov 09, 2022
TensorFlow Tutorials with YouTube Videos

TensorFlow Tutorials Original repository on GitHub Original author is Magnus Erik Hvass Pedersen Introduction These tutorials are intended for beginne

9.1k Jan 02, 2023
The Rich Get Richer: Disparate Impact of Semi-Supervised Learning

The Rich Get Richer: Disparate Impact of Semi-Supervised Learning Preprocess file of the dataset used in implicit sub-populations: (Demographic groups

<a href=[email protected]"> 4 Oct 14, 2022
(AAAI2020)Grapy-ML: Graph Pyramid Mutual Learning for Cross-dataset Human Parsing

Grapy-ML: Graph Pyramid Mutual Learning for Cross-dataset Human Parsing This repository contains pytorch source code for AAAI2020 oral paper: Grapy-ML

54 Aug 04, 2022
Code for our SIGCOMM'21 paper "Network Planning with Deep Reinforcement Learning".

0. Introduction This repository contains the source code for our SIGCOMM'21 paper "Network Planning with Deep Reinforcement Learning". Notes The netwo

NetX Group 68 Nov 24, 2022
RL and distillation in CARLA using a factorized world model

World on Rails Learning to drive from a world on rails Dian Chen, Vladlen Koltun, Philipp Krähenbühl, arXiv techical report (arXiv 2105.00636) This re

Dian Chen 131 Dec 16, 2022
Sdf sparse conv - Deep Learning on SDF for Classifying Brain Biomarkers

Deep Learning on SDF for Classifying Brain Biomarkers To reproduce the results f

1 Jan 25, 2022